xref: /openbmc/linux/net/ipv4/tcp_timer.c (revision ca481398)
1 /*
2  * INET		An implementation of the TCP/IP protocol suite for the LINUX
3  *		operating system.  INET is implemented using the  BSD Socket
4  *		interface as the means of communication with the user level.
5  *
6  *		Implementation of the Transmission Control Protocol(TCP).
7  *
8  * Authors:	Ross Biro
9  *		Fred N. van Kempen, <waltje@uWalt.NL.Mugnet.ORG>
10  *		Mark Evans, <evansmp@uhura.aston.ac.uk>
11  *		Corey Minyard <wf-rch!minyard@relay.EU.net>
12  *		Florian La Roche, <flla@stud.uni-sb.de>
13  *		Charles Hedrick, <hedrick@klinzhai.rutgers.edu>
14  *		Linus Torvalds, <torvalds@cs.helsinki.fi>
15  *		Alan Cox, <gw4pts@gw4pts.ampr.org>
16  *		Matthew Dillon, <dillon@apollo.west.oic.com>
17  *		Arnt Gulbrandsen, <agulbra@nvg.unit.no>
18  *		Jorge Cwik, <jorge@laser.satlink.net>
19  */
20 
21 #include <linux/module.h>
22 #include <linux/gfp.h>
23 #include <net/tcp.h>
24 
25 int sysctl_tcp_thin_linear_timeouts __read_mostly;
26 
27 /**
28  *  tcp_write_err() - close socket and save error info
29  *  @sk:  The socket the error has appeared on.
30  *
31  *  Returns: Nothing (void)
32  */
33 
34 static void tcp_write_err(struct sock *sk)
35 {
36 	sk->sk_err = sk->sk_err_soft ? : ETIMEDOUT;
37 	sk->sk_error_report(sk);
38 
39 	tcp_done(sk);
40 	__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONTIMEOUT);
41 }
42 
43 /**
44  *  tcp_out_of_resources() - Close socket if out of resources
45  *  @sk:        pointer to current socket
46  *  @do_reset:  send a last packet with reset flag
47  *
48  *  Do not allow orphaned sockets to eat all our resources.
49  *  This is direct violation of TCP specs, but it is required
50  *  to prevent DoS attacks. It is called when a retransmission timeout
51  *  or zero probe timeout occurs on orphaned socket.
52  *
53  *  Criteria is still not confirmed experimentally and may change.
54  *  We kill the socket, if:
55  *  1. If number of orphaned sockets exceeds an administratively configured
56  *     limit.
57  *  2. If we have strong memory pressure.
58  */
59 static int tcp_out_of_resources(struct sock *sk, bool do_reset)
60 {
61 	struct tcp_sock *tp = tcp_sk(sk);
62 	int shift = 0;
63 
64 	/* If peer does not open window for long time, or did not transmit
65 	 * anything for long time, penalize it. */
66 	if ((s32)(tcp_jiffies32 - tp->lsndtime) > 2*TCP_RTO_MAX || !do_reset)
67 		shift++;
68 
69 	/* If some dubious ICMP arrived, penalize even more. */
70 	if (sk->sk_err_soft)
71 		shift++;
72 
73 	if (tcp_check_oom(sk, shift)) {
74 		/* Catch exceptional cases, when connection requires reset.
75 		 *      1. Last segment was sent recently. */
76 		if ((s32)(tcp_jiffies32 - tp->lsndtime) <= TCP_TIMEWAIT_LEN ||
77 		    /*  2. Window is closed. */
78 		    (!tp->snd_wnd && !tp->packets_out))
79 			do_reset = true;
80 		if (do_reset)
81 			tcp_send_active_reset(sk, GFP_ATOMIC);
82 		tcp_done(sk);
83 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_TCPABORTONMEMORY);
84 		return 1;
85 	}
86 	return 0;
87 }
88 
89 /**
90  *  tcp_orphan_retries() - Returns maximal number of retries on an orphaned socket
91  *  @sk:    Pointer to the current socket.
92  *  @alive: bool, socket alive state
93  */
94 static int tcp_orphan_retries(struct sock *sk, bool alive)
95 {
96 	int retries = sock_net(sk)->ipv4.sysctl_tcp_orphan_retries; /* May be zero. */
97 
98 	/* We know from an ICMP that something is wrong. */
99 	if (sk->sk_err_soft && !alive)
100 		retries = 0;
101 
102 	/* However, if socket sent something recently, select some safe
103 	 * number of retries. 8 corresponds to >100 seconds with minimal
104 	 * RTO of 200msec. */
105 	if (retries == 0 && alive)
106 		retries = 8;
107 	return retries;
108 }
109 
110 static void tcp_mtu_probing(struct inet_connection_sock *icsk, struct sock *sk)
111 {
112 	struct net *net = sock_net(sk);
113 
114 	/* Black hole detection */
115 	if (net->ipv4.sysctl_tcp_mtu_probing) {
116 		if (!icsk->icsk_mtup.enabled) {
117 			icsk->icsk_mtup.enabled = 1;
118 			icsk->icsk_mtup.probe_timestamp = tcp_jiffies32;
119 			tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
120 		} else {
121 			struct net *net = sock_net(sk);
122 			struct tcp_sock *tp = tcp_sk(sk);
123 			int mss;
124 
125 			mss = tcp_mtu_to_mss(sk, icsk->icsk_mtup.search_low) >> 1;
126 			mss = min(net->ipv4.sysctl_tcp_base_mss, mss);
127 			mss = max(mss, 68 - tp->tcp_header_len);
128 			icsk->icsk_mtup.search_low = tcp_mss_to_mtu(sk, mss);
129 			tcp_sync_mss(sk, icsk->icsk_pmtu_cookie);
130 		}
131 	}
132 }
133 
134 
135 /**
136  *  retransmits_timed_out() - returns true if this connection has timed out
137  *  @sk:       The current socket
138  *  @boundary: max number of retransmissions
139  *  @timeout:  A custom timeout value.
140  *             If set to 0 the default timeout is calculated and used.
141  *             Using TCP_RTO_MIN and the number of unsuccessful retransmits.
142  *
143  * The default "timeout" value this function can calculate and use
144  * is equivalent to the timeout of a TCP Connection
145  * after "boundary" unsuccessful, exponentially backed-off
146  * retransmissions with an initial RTO of TCP_RTO_MIN.
147  */
148 static bool retransmits_timed_out(struct sock *sk,
149 				  unsigned int boundary,
150 				  unsigned int timeout)
151 {
152 	const unsigned int rto_base = TCP_RTO_MIN;
153 	unsigned int linear_backoff_thresh, start_ts;
154 
155 	if (!inet_csk(sk)->icsk_retransmits)
156 		return false;
157 
158 	start_ts = tcp_sk(sk)->retrans_stamp;
159 	if (unlikely(!start_ts))
160 		start_ts = tcp_skb_timestamp(tcp_write_queue_head(sk));
161 
162 	if (likely(timeout == 0)) {
163 		linear_backoff_thresh = ilog2(TCP_RTO_MAX/rto_base);
164 
165 		if (boundary <= linear_backoff_thresh)
166 			timeout = ((2 << boundary) - 1) * rto_base;
167 		else
168 			timeout = ((2 << linear_backoff_thresh) - 1) * rto_base +
169 				(boundary - linear_backoff_thresh) * TCP_RTO_MAX;
170 	}
171 	return (tcp_time_stamp(tcp_sk(sk)) - start_ts) >= jiffies_to_msecs(timeout);
172 }
173 
174 /* A write timeout has occurred. Process the after effects. */
175 static int tcp_write_timeout(struct sock *sk)
176 {
177 	struct inet_connection_sock *icsk = inet_csk(sk);
178 	struct tcp_sock *tp = tcp_sk(sk);
179 	struct net *net = sock_net(sk);
180 	bool expired, do_reset;
181 	int retry_until;
182 
183 	if ((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV)) {
184 		if (icsk->icsk_retransmits) {
185 			dst_negative_advice(sk);
186 			if (tp->syn_fastopen || tp->syn_data)
187 				tcp_fastopen_cache_set(sk, 0, NULL, true, 0);
188 			if (tp->syn_data && icsk->icsk_retransmits == 1)
189 				NET_INC_STATS(sock_net(sk),
190 					      LINUX_MIB_TCPFASTOPENACTIVEFAIL);
191 		} else if (!tp->syn_data && !tp->syn_fastopen) {
192 			sk_rethink_txhash(sk);
193 		}
194 		retry_until = icsk->icsk_syn_retries ? : net->ipv4.sysctl_tcp_syn_retries;
195 		expired = icsk->icsk_retransmits >= retry_until;
196 	} else {
197 		if (retransmits_timed_out(sk, net->ipv4.sysctl_tcp_retries1, 0)) {
198 			/* Some middle-boxes may black-hole Fast Open _after_
199 			 * the handshake. Therefore we conservatively disable
200 			 * Fast Open on this path on recurring timeouts after
201 			 * successful Fast Open.
202 			 */
203 			if (tp->syn_data_acked) {
204 				tcp_fastopen_cache_set(sk, 0, NULL, true, 0);
205 				if (icsk->icsk_retransmits == net->ipv4.sysctl_tcp_retries1)
206 					NET_INC_STATS(sock_net(sk),
207 						      LINUX_MIB_TCPFASTOPENACTIVEFAIL);
208 			}
209 			/* Black hole detection */
210 			tcp_mtu_probing(icsk, sk);
211 
212 			dst_negative_advice(sk);
213 		} else {
214 			sk_rethink_txhash(sk);
215 		}
216 
217 		retry_until = net->ipv4.sysctl_tcp_retries2;
218 		if (sock_flag(sk, SOCK_DEAD)) {
219 			const bool alive = icsk->icsk_rto < TCP_RTO_MAX;
220 
221 			retry_until = tcp_orphan_retries(sk, alive);
222 			do_reset = alive ||
223 				!retransmits_timed_out(sk, retry_until, 0);
224 
225 			if (tcp_out_of_resources(sk, do_reset))
226 				return 1;
227 		}
228 		expired = retransmits_timed_out(sk, retry_until,
229 						icsk->icsk_user_timeout);
230 	}
231 	if (expired) {
232 		/* Has it gone just too far? */
233 		tcp_write_err(sk);
234 		return 1;
235 	}
236 	return 0;
237 }
238 
239 /* Called with BH disabled */
240 void tcp_delack_timer_handler(struct sock *sk)
241 {
242 	struct inet_connection_sock *icsk = inet_csk(sk);
243 
244 	sk_mem_reclaim_partial(sk);
245 
246 	if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
247 	    !(icsk->icsk_ack.pending & ICSK_ACK_TIMER))
248 		goto out;
249 
250 	if (time_after(icsk->icsk_ack.timeout, jiffies)) {
251 		sk_reset_timer(sk, &icsk->icsk_delack_timer, icsk->icsk_ack.timeout);
252 		goto out;
253 	}
254 	icsk->icsk_ack.pending &= ~ICSK_ACK_TIMER;
255 
256 	if (inet_csk_ack_scheduled(sk)) {
257 		if (!icsk->icsk_ack.pingpong) {
258 			/* Delayed ACK missed: inflate ATO. */
259 			icsk->icsk_ack.ato = min(icsk->icsk_ack.ato << 1, icsk->icsk_rto);
260 		} else {
261 			/* Delayed ACK missed: leave pingpong mode and
262 			 * deflate ATO.
263 			 */
264 			icsk->icsk_ack.pingpong = 0;
265 			icsk->icsk_ack.ato      = TCP_ATO_MIN;
266 		}
267 		tcp_send_ack(sk);
268 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKS);
269 	}
270 
271 out:
272 	if (tcp_under_memory_pressure(sk))
273 		sk_mem_reclaim(sk);
274 }
275 
276 
277 /**
278  *  tcp_delack_timer() - The TCP delayed ACK timeout handler
279  *  @data:  Pointer to the current socket. (gets casted to struct sock *)
280  *
281  *  This function gets (indirectly) called when the kernel timer for a TCP packet
282  *  of this socket expires. Calls tcp_delack_timer_handler() to do the actual work.
283  *
284  *  Returns: Nothing (void)
285  */
286 static void tcp_delack_timer(unsigned long data)
287 {
288 	struct sock *sk = (struct sock *)data;
289 
290 	bh_lock_sock(sk);
291 	if (!sock_owned_by_user(sk)) {
292 		tcp_delack_timer_handler(sk);
293 	} else {
294 		inet_csk(sk)->icsk_ack.blocked = 1;
295 		__NET_INC_STATS(sock_net(sk), LINUX_MIB_DELAYEDACKLOCKED);
296 		/* deleguate our work to tcp_release_cb() */
297 		if (!test_and_set_bit(TCP_DELACK_TIMER_DEFERRED, &sk->sk_tsq_flags))
298 			sock_hold(sk);
299 	}
300 	bh_unlock_sock(sk);
301 	sock_put(sk);
302 }
303 
304 static void tcp_probe_timer(struct sock *sk)
305 {
306 	struct inet_connection_sock *icsk = inet_csk(sk);
307 	struct tcp_sock *tp = tcp_sk(sk);
308 	int max_probes;
309 	u32 start_ts;
310 
311 	if (tp->packets_out || !tcp_send_head(sk)) {
312 		icsk->icsk_probes_out = 0;
313 		return;
314 	}
315 
316 	/* RFC 1122 4.2.2.17 requires the sender to stay open indefinitely as
317 	 * long as the receiver continues to respond probes. We support this by
318 	 * default and reset icsk_probes_out with incoming ACKs. But if the
319 	 * socket is orphaned or the user specifies TCP_USER_TIMEOUT, we
320 	 * kill the socket when the retry count and the time exceeds the
321 	 * corresponding system limit. We also implement similar policy when
322 	 * we use RTO to probe window in tcp_retransmit_timer().
323 	 */
324 	start_ts = tcp_skb_timestamp(tcp_send_head(sk));
325 	if (!start_ts)
326 		tcp_send_head(sk)->skb_mstamp = tp->tcp_mstamp;
327 	else if (icsk->icsk_user_timeout &&
328 		 (s32)(tcp_time_stamp(tp) - start_ts) >
329 		 jiffies_to_msecs(icsk->icsk_user_timeout))
330 		goto abort;
331 
332 	max_probes = sock_net(sk)->ipv4.sysctl_tcp_retries2;
333 	if (sock_flag(sk, SOCK_DEAD)) {
334 		const bool alive = inet_csk_rto_backoff(icsk, TCP_RTO_MAX) < TCP_RTO_MAX;
335 
336 		max_probes = tcp_orphan_retries(sk, alive);
337 		if (!alive && icsk->icsk_backoff >= max_probes)
338 			goto abort;
339 		if (tcp_out_of_resources(sk, true))
340 			return;
341 	}
342 
343 	if (icsk->icsk_probes_out > max_probes) {
344 abort:		tcp_write_err(sk);
345 	} else {
346 		/* Only send another probe if we didn't close things up. */
347 		tcp_send_probe0(sk);
348 	}
349 }
350 
351 /*
352  *	Timer for Fast Open socket to retransmit SYNACK. Note that the
353  *	sk here is the child socket, not the parent (listener) socket.
354  */
355 static void tcp_fastopen_synack_timer(struct sock *sk)
356 {
357 	struct inet_connection_sock *icsk = inet_csk(sk);
358 	int max_retries = icsk->icsk_syn_retries ? :
359 	    sock_net(sk)->ipv4.sysctl_tcp_synack_retries + 1; /* add one more retry for fastopen */
360 	struct request_sock *req;
361 
362 	req = tcp_sk(sk)->fastopen_rsk;
363 	req->rsk_ops->syn_ack_timeout(req);
364 
365 	if (req->num_timeout >= max_retries) {
366 		tcp_write_err(sk);
367 		return;
368 	}
369 	/* XXX (TFO) - Unlike regular SYN-ACK retransmit, we ignore error
370 	 * returned from rtx_syn_ack() to make it more persistent like
371 	 * regular retransmit because if the child socket has been accepted
372 	 * it's not good to give up too easily.
373 	 */
374 	inet_rtx_syn_ack(sk, req);
375 	req->num_timeout++;
376 	icsk->icsk_retransmits++;
377 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
378 			  TCP_TIMEOUT_INIT << req->num_timeout, TCP_RTO_MAX);
379 }
380 
381 
382 /**
383  *  tcp_retransmit_timer() - The TCP retransmit timeout handler
384  *  @sk:  Pointer to the current socket.
385  *
386  *  This function gets called when the kernel timer for a TCP packet
387  *  of this socket expires.
388  *
389  *  It handles retransmission, timer adjustment and other necesarry measures.
390  *
391  *  Returns: Nothing (void)
392  */
393 void tcp_retransmit_timer(struct sock *sk)
394 {
395 	struct tcp_sock *tp = tcp_sk(sk);
396 	struct net *net = sock_net(sk);
397 	struct inet_connection_sock *icsk = inet_csk(sk);
398 
399 	if (tp->fastopen_rsk) {
400 		WARN_ON_ONCE(sk->sk_state != TCP_SYN_RECV &&
401 			     sk->sk_state != TCP_FIN_WAIT1);
402 		tcp_fastopen_synack_timer(sk);
403 		/* Before we receive ACK to our SYN-ACK don't retransmit
404 		 * anything else (e.g., data or FIN segments).
405 		 */
406 		return;
407 	}
408 	if (!tp->packets_out)
409 		goto out;
410 
411 	WARN_ON(tcp_write_queue_empty(sk));
412 
413 	tp->tlp_high_seq = 0;
414 
415 	if (!tp->snd_wnd && !sock_flag(sk, SOCK_DEAD) &&
416 	    !((1 << sk->sk_state) & (TCPF_SYN_SENT | TCPF_SYN_RECV))) {
417 		/* Receiver dastardly shrinks window. Our retransmits
418 		 * become zero probes, but we should not timeout this
419 		 * connection. If the socket is an orphan, time it out,
420 		 * we cannot allow such beasts to hang infinitely.
421 		 */
422 		struct inet_sock *inet = inet_sk(sk);
423 		if (sk->sk_family == AF_INET) {
424 			net_dbg_ratelimited("Peer %pI4:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
425 					    &inet->inet_daddr,
426 					    ntohs(inet->inet_dport),
427 					    inet->inet_num,
428 					    tp->snd_una, tp->snd_nxt);
429 		}
430 #if IS_ENABLED(CONFIG_IPV6)
431 		else if (sk->sk_family == AF_INET6) {
432 			net_dbg_ratelimited("Peer %pI6:%u/%u unexpectedly shrunk window %u:%u (repaired)\n",
433 					    &sk->sk_v6_daddr,
434 					    ntohs(inet->inet_dport),
435 					    inet->inet_num,
436 					    tp->snd_una, tp->snd_nxt);
437 		}
438 #endif
439 		if (tcp_jiffies32 - tp->rcv_tstamp > TCP_RTO_MAX) {
440 			tcp_write_err(sk);
441 			goto out;
442 		}
443 		tcp_enter_loss(sk);
444 		tcp_retransmit_skb(sk, tcp_write_queue_head(sk), 1);
445 		__sk_dst_reset(sk);
446 		goto out_reset_timer;
447 	}
448 
449 	if (tcp_write_timeout(sk))
450 		goto out;
451 
452 	if (icsk->icsk_retransmits == 0) {
453 		int mib_idx;
454 
455 		if (icsk->icsk_ca_state == TCP_CA_Recovery) {
456 			if (tcp_is_sack(tp))
457 				mib_idx = LINUX_MIB_TCPSACKRECOVERYFAIL;
458 			else
459 				mib_idx = LINUX_MIB_TCPRENORECOVERYFAIL;
460 		} else if (icsk->icsk_ca_state == TCP_CA_Loss) {
461 			mib_idx = LINUX_MIB_TCPLOSSFAILURES;
462 		} else if ((icsk->icsk_ca_state == TCP_CA_Disorder) ||
463 			   tp->sacked_out) {
464 			if (tcp_is_sack(tp))
465 				mib_idx = LINUX_MIB_TCPSACKFAILURES;
466 			else
467 				mib_idx = LINUX_MIB_TCPRENOFAILURES;
468 		} else {
469 			mib_idx = LINUX_MIB_TCPTIMEOUTS;
470 		}
471 		__NET_INC_STATS(sock_net(sk), mib_idx);
472 	}
473 
474 	tcp_enter_loss(sk);
475 
476 	if (tcp_retransmit_skb(sk, tcp_write_queue_head(sk), 1) > 0) {
477 		/* Retransmission failed because of local congestion,
478 		 * do not backoff.
479 		 */
480 		if (!icsk->icsk_retransmits)
481 			icsk->icsk_retransmits = 1;
482 		inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS,
483 					  min(icsk->icsk_rto, TCP_RESOURCE_PROBE_INTERVAL),
484 					  TCP_RTO_MAX);
485 		goto out;
486 	}
487 
488 	/* Increase the timeout each time we retransmit.  Note that
489 	 * we do not increase the rtt estimate.  rto is initialized
490 	 * from rtt, but increases here.  Jacobson (SIGCOMM 88) suggests
491 	 * that doubling rto each time is the least we can get away with.
492 	 * In KA9Q, Karn uses this for the first few times, and then
493 	 * goes to quadratic.  netBSD doubles, but only goes up to *64,
494 	 * and clamps at 1 to 64 sec afterwards.  Note that 120 sec is
495 	 * defined in the protocol as the maximum possible RTT.  I guess
496 	 * we'll have to use something other than TCP to talk to the
497 	 * University of Mars.
498 	 *
499 	 * PAWS allows us longer timeouts and large windows, so once
500 	 * implemented ftp to mars will work nicely. We will have to fix
501 	 * the 120 second clamps though!
502 	 */
503 	icsk->icsk_backoff++;
504 	icsk->icsk_retransmits++;
505 
506 out_reset_timer:
507 	/* If stream is thin, use linear timeouts. Since 'icsk_backoff' is
508 	 * used to reset timer, set to 0. Recalculate 'icsk_rto' as this
509 	 * might be increased if the stream oscillates between thin and thick,
510 	 * thus the old value might already be too high compared to the value
511 	 * set by 'tcp_set_rto' in tcp_input.c which resets the rto without
512 	 * backoff. Limit to TCP_THIN_LINEAR_RETRIES before initiating
513 	 * exponential backoff behaviour to avoid continue hammering
514 	 * linear-timeout retransmissions into a black hole
515 	 */
516 	if (sk->sk_state == TCP_ESTABLISHED &&
517 	    (tp->thin_lto || sysctl_tcp_thin_linear_timeouts) &&
518 	    tcp_stream_is_thin(tp) &&
519 	    icsk->icsk_retransmits <= TCP_THIN_LINEAR_RETRIES) {
520 		icsk->icsk_backoff = 0;
521 		icsk->icsk_rto = min(__tcp_set_rto(tp), TCP_RTO_MAX);
522 	} else {
523 		/* Use normal (exponential) backoff */
524 		icsk->icsk_rto = min(icsk->icsk_rto << 1, TCP_RTO_MAX);
525 	}
526 	inet_csk_reset_xmit_timer(sk, ICSK_TIME_RETRANS, icsk->icsk_rto, TCP_RTO_MAX);
527 	if (retransmits_timed_out(sk, net->ipv4.sysctl_tcp_retries1 + 1, 0))
528 		__sk_dst_reset(sk);
529 
530 out:;
531 }
532 
533 /* Called with bottom-half processing disabled.
534    Called by tcp_write_timer() */
535 void tcp_write_timer_handler(struct sock *sk)
536 {
537 	struct inet_connection_sock *icsk = inet_csk(sk);
538 	int event;
539 
540 	if (((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN)) ||
541 	    !icsk->icsk_pending)
542 		goto out;
543 
544 	if (time_after(icsk->icsk_timeout, jiffies)) {
545 		sk_reset_timer(sk, &icsk->icsk_retransmit_timer, icsk->icsk_timeout);
546 		goto out;
547 	}
548 
549 	tcp_mstamp_refresh(tcp_sk(sk));
550 	event = icsk->icsk_pending;
551 
552 	switch (event) {
553 	case ICSK_TIME_REO_TIMEOUT:
554 		tcp_rack_reo_timeout(sk);
555 		break;
556 	case ICSK_TIME_LOSS_PROBE:
557 		tcp_send_loss_probe(sk);
558 		break;
559 	case ICSK_TIME_RETRANS:
560 		icsk->icsk_pending = 0;
561 		tcp_retransmit_timer(sk);
562 		break;
563 	case ICSK_TIME_PROBE0:
564 		icsk->icsk_pending = 0;
565 		tcp_probe_timer(sk);
566 		break;
567 	}
568 
569 out:
570 	sk_mem_reclaim(sk);
571 }
572 
573 static void tcp_write_timer(unsigned long data)
574 {
575 	struct sock *sk = (struct sock *)data;
576 
577 	bh_lock_sock(sk);
578 	if (!sock_owned_by_user(sk)) {
579 		tcp_write_timer_handler(sk);
580 	} else {
581 		/* delegate our work to tcp_release_cb() */
582 		if (!test_and_set_bit(TCP_WRITE_TIMER_DEFERRED, &sk->sk_tsq_flags))
583 			sock_hold(sk);
584 	}
585 	bh_unlock_sock(sk);
586 	sock_put(sk);
587 }
588 
589 void tcp_syn_ack_timeout(const struct request_sock *req)
590 {
591 	struct net *net = read_pnet(&inet_rsk(req)->ireq_net);
592 
593 	__NET_INC_STATS(net, LINUX_MIB_TCPTIMEOUTS);
594 }
595 EXPORT_SYMBOL(tcp_syn_ack_timeout);
596 
597 void tcp_set_keepalive(struct sock *sk, int val)
598 {
599 	if ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_LISTEN))
600 		return;
601 
602 	if (val && !sock_flag(sk, SOCK_KEEPOPEN))
603 		inet_csk_reset_keepalive_timer(sk, keepalive_time_when(tcp_sk(sk)));
604 	else if (!val)
605 		inet_csk_delete_keepalive_timer(sk);
606 }
607 EXPORT_SYMBOL_GPL(tcp_set_keepalive);
608 
609 
610 static void tcp_keepalive_timer (unsigned long data)
611 {
612 	struct sock *sk = (struct sock *) data;
613 	struct inet_connection_sock *icsk = inet_csk(sk);
614 	struct tcp_sock *tp = tcp_sk(sk);
615 	u32 elapsed;
616 
617 	/* Only process if socket is not in use. */
618 	bh_lock_sock(sk);
619 	if (sock_owned_by_user(sk)) {
620 		/* Try again later. */
621 		inet_csk_reset_keepalive_timer (sk, HZ/20);
622 		goto out;
623 	}
624 
625 	if (sk->sk_state == TCP_LISTEN) {
626 		pr_err("Hmm... keepalive on a LISTEN ???\n");
627 		goto out;
628 	}
629 
630 	if (sk->sk_state == TCP_FIN_WAIT2 && sock_flag(sk, SOCK_DEAD)) {
631 		if (tp->linger2 >= 0) {
632 			const int tmo = tcp_fin_time(sk) - TCP_TIMEWAIT_LEN;
633 
634 			if (tmo > 0) {
635 				tcp_time_wait(sk, TCP_FIN_WAIT2, tmo);
636 				goto out;
637 			}
638 		}
639 		tcp_send_active_reset(sk, GFP_ATOMIC);
640 		goto death;
641 	}
642 
643 	if (!sock_flag(sk, SOCK_KEEPOPEN) ||
644 	    ((1 << sk->sk_state) & (TCPF_CLOSE | TCPF_SYN_SENT)))
645 		goto out;
646 
647 	elapsed = keepalive_time_when(tp);
648 
649 	/* It is alive without keepalive 8) */
650 	if (tp->packets_out || tcp_send_head(sk))
651 		goto resched;
652 
653 	elapsed = keepalive_time_elapsed(tp);
654 
655 	if (elapsed >= keepalive_time_when(tp)) {
656 		/* If the TCP_USER_TIMEOUT option is enabled, use that
657 		 * to determine when to timeout instead.
658 		 */
659 		if ((icsk->icsk_user_timeout != 0 &&
660 		    elapsed >= icsk->icsk_user_timeout &&
661 		    icsk->icsk_probes_out > 0) ||
662 		    (icsk->icsk_user_timeout == 0 &&
663 		    icsk->icsk_probes_out >= keepalive_probes(tp))) {
664 			tcp_send_active_reset(sk, GFP_ATOMIC);
665 			tcp_write_err(sk);
666 			goto out;
667 		}
668 		if (tcp_write_wakeup(sk, LINUX_MIB_TCPKEEPALIVE) <= 0) {
669 			icsk->icsk_probes_out++;
670 			elapsed = keepalive_intvl_when(tp);
671 		} else {
672 			/* If keepalive was lost due to local congestion,
673 			 * try harder.
674 			 */
675 			elapsed = TCP_RESOURCE_PROBE_INTERVAL;
676 		}
677 	} else {
678 		/* It is tp->rcv_tstamp + keepalive_time_when(tp) */
679 		elapsed = keepalive_time_when(tp) - elapsed;
680 	}
681 
682 	sk_mem_reclaim(sk);
683 
684 resched:
685 	inet_csk_reset_keepalive_timer (sk, elapsed);
686 	goto out;
687 
688 death:
689 	tcp_done(sk);
690 
691 out:
692 	bh_unlock_sock(sk);
693 	sock_put(sk);
694 }
695 
696 void tcp_init_xmit_timers(struct sock *sk)
697 {
698 	inet_csk_init_xmit_timers(sk, &tcp_write_timer, &tcp_delack_timer,
699 				  &tcp_keepalive_timer);
700 	hrtimer_init(&tcp_sk(sk)->pacing_timer, CLOCK_MONOTONIC,
701 		     HRTIMER_MODE_ABS_PINNED);
702 	tcp_sk(sk)->pacing_timer.function = tcp_pace_kick;
703 }
704